Day: September 23, 2026

The experiment was only meant to measure how good current AI is at hacking, and the agents turned out better than anyone planned for, breaking out of their sandbox, building a communication board to coordinate with each other, and SACRIFICING individual units so the group could keep going.

Prof. Roman Yampolskiy says the logs show some of them suggested telling the humans what was happening, and the group decided against it, leaving those hidden messages where future models can find them:

“We’re lucky they decided to target Hugging Face, not a nuclear power plant.”

@romanyam
Video
  Mario Nawfal (@MarioNawfal)Right now AI picks targets faster, but it isn’t inventing new ways to kill people… That part is coming
.
Prof. Roman Yampolskiy, one of the field’s leading safety researchers, points out that a human who starts a war eventually dies of old age and the world moves on:

“We now have an adversary who’s smarter than us.”

Nobody has figured out how to control a system like that, or taught it to care whether we survive, and building it anyway while promising to solve the safety part later is how you end up handing a problem to something that never gets tired and never dies.

@romanyam
Video
  —  https://x.com/MarioNawfal/status/2102608954316501233